A Foldit plugin for running RC Foundry machine-learning models — RoseTTAFold3, RFdiffusion3, and LigandMPNN — as an in-game backend.
It brings three AI operations to the app:
| Button | Op | Model | What it does |
|---|---|---|---|
| Predict | rf3_predict |
RoseTTAFold3 | Re-predict the structure of the focused entity (or the whole structure when nothing is focused) |
| Design | rfd3_design |
RFdiffusion3 | Generate a binder against the focused protein or ligand |
| MPNN | mpnn_design |
LigandMPNN | Redesign the focused chain's sequence, holding the selected residues fixed |
It also exposes a sequence_design query that returns scored sequence
candidates for a future UI panel.
This is a kind = "python" plugin. Foldit hosts it out-of-process through
foldit-python-host; the entry module is foundry_plugin. The heavy model code
(the foundry package) lives under deps/foundry and is imported by the thin
plugin wrapper.
Model weights are large and are not committed — they live under
assets/weights/rc_foundry/ and are resolved at load time. On Apple hardware
the models run on the MPS (Metal) GPU backend.
Python plugins use pixi for their environment. Set the plugin up (create the env, resolve weights) through the workspace xtask:
cargo xtask setup-plugins foundryThe plugin subclasses the PluginInterface base class from the Foldit plugin
SDK; see the Foldit workspace docs ("Python and Native Plugins") for how the
host discovers and loads a Python plugin.